Detection of SST fronts from high-resolution model - PowerPoint Presentation

Slide1

Detection of SST fronts from high-resolution model hindcast results and its preliminary evaluation in the South China Sea

Shihe Rena, Xueming Zhua, and Drevillon Marieba National Marine Environmental Forcasting Center, Beijing, Chinab Mercator Océan, Ramonville Saint Agne, France

EGU, 8 May 2020

Slide2

OutlineBackgroundDataFront detection algorithmResults and discussionConclusion

Slide3

front detection methods gradient-based approach with Canny operator (Canny, 1986; Kostianoy et al., 2004; Breaker et al., 2005; Belkin and Oreilly 2009; Oram et al. 2008; Kirches et al. 2016; ......)histogram-based approach: single edge detection algorithm (SIED)

(Cayula and Cornillon 1992; Miller, 2004&2009; Ullman and Cornillon, 1999; Hickox et al., 2000;Nieto and Demarcq,2006; ......)other methodsstatistical approach (Ping et al. 2015; Hopkins et al. 2010; ..... )entropic approach (Shimada et al., 2005;.....)Convolutional Neural Network (Lima et al. 2017;....)

.......

Slide4

Automated front detection tools for HR-model resultsDetection for HR-model data is different with satellite data.Specific pre-processing and post-processing procedure are needed to ensure balance of frontal continuity and positioning accuracy.We need highlight the quasi-stationary frontal position and

primary/strong fronts Frontal parameters and products are needed.The aim is to build mesoscale frontal forecasting system with detection, forecasting and validation.

Slide5

OutlineBackgroundDataFront detection algorithmResults Conclusion

Slide6

long-term daily SST(2008-2017)regional ROMS daily average hindcast in the South China Sea (1/30o)OSTIA daily SST (5km)

Mercator (1/12o) global reanalysis (PSY4V3)

Slide7

OutlineBackgroundDataFront detection algorithmResults Conclusion

Slide8

front detection algorithm

Frontal line Frontal Zone

detected result snapshot (frontal length greater than 100km)

Slide9

frontal line detectionfront edge localization with Canny methodfront

edge following to deal with branch pixelsfront edge merging to link nearby segments

Slide10

Canny edge detection stepsCalculate front gradient and direction Non-maximum suppression

determine local maxima along gradient directionchange non-maximum pixels value to zeroHysteresis thresholding Large threshold

:

less edge points, more gaps

Small threshold

:

more edge points, more false edges

E

dge linking

Strong edge

Weak edges

Weak edge removed

edges

Canny, 1986, A computational approach to edge detection, IEEE Transactions on Pattern Analysis and Machine Intelligence 8(6) , pp. 679–698

Slide11

front edge followingbranch pixels still exist after edge localization (9 junctions occur in test data)necessary to extract and find which edge should be followed based on front direction and magnitude

separate branches to examine the statistics of each individual front line

frontal line extracted after edge localization

Slide12

front edge following

start pixel

center pixel

non-front pixel

along-front pixel

2 pixels

1 pixels

0

pixels

Slide13

front edge mergingafter merging, frontline number from 166 to 117

before merge

after merge

(a)

(b)

(d)

(c)

Slide14

OutlineBackgroundDataFront detection algorithmResults Conclusion

Slide15

Climatology: SST gradientOSTIA

ROMSMercator

Slide16

climatology: front area occurence frequency

OSTIAROMSMercator

Slide17

climatology: front line occurence frequency

OSTIAROMSMercator

Slide18

OSTIA

ROMS

Mercator

Slide19

front area occurence frequency

OSTIAROMSMercator

Slide20

ROMS - OSTIA

Mercator - OSTIA

front area occurence frequency(bias)

Slide21

climatology of front parameter

width

area

number

strength

length

seasonal varibility

Slide22

summarydeveloping frontal detecting toolsfinish a stable frontal detection tools performing frontal statistical with different SST datasetsdaily, monthly, seasonal ....

frontal parameter, occurence frequency

Slide23

Thank you for attention